fix(ui): group cost optimization cache leakage by model group (#43008)

* test(ui): cover cost optimization cache leakage grouping by model group

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(ui): group cost optimization cache leakage by model group

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

tests/e2e/ui/fixtures/pages.ts

@@ -20,6 +20,7 @@ export enum Page {
   RouterSettings = "router-settings",
   UiTheme = "ui-theme",
   CostTracking = "cost-tracking",
+  CostOptimization = "cost-optimization",
   ModelHubTable = "model-hub-table",
   Caching = "caching",
   Logs = "logs",

tests/e2e/ui/tests/integrationCritical/costOptimizationModelGroups.spec.ts

@@ -0,0 +1,51 @@
+import { test, expect } from "@playwright/test";
+import { randomUUID } from "node:crypto";
+import { execFileSync } from "node:child_process";
+import * as path from "node:path";
+import { Page } from "../../fixtures/pages";
+import { navigateToPage } from "../../helpers/navigation";
+
+test("cache leakage by model merges a deployment's resolved and requested model names into its model group", async ({
+  page,
+}) => {
+  const master = process.env.LITELLM_MASTER_KEY ?? "sk-integration-master";
+  const marker = `integration-browser-${randomUUID()}`;
+  const group = `${marker}-public`;
+  const deployment = `${marker}-backend`;
+  const apiKey = marker;
+  const support = (...args: string[]) =>
+    execFileSync(
+      process.env.INTEGRATION_PYTHON ?? "python",
+      [
+        path.resolve(
+          __dirname,
+          "../../../../integration/_support/daily_spend_rows.py",
+        ),
+        ...args,
+      ],
+      { encoding: "utf8", timeout: 10_000, killSignal: "SIGKILL" },
+    );
+  try {
+    support("seed", apiKey, group, deployment);
+    await page.goto("/ui/login");
+    await page.getByPlaceholder("Enter your username").fill("admin");
+    await page.getByPlaceholder("Enter your password").fill(master);
+    await page.getByRole("button", { name: "Login", exact: true }).click();
+    await expect(page).toHaveURL(
+      (url) =>
+        url.pathname.startsWith("/ui") && !url.pathname.includes("login"),
+    );
+    await navigateToPage(page, Page.CostOptimization);
+    await page.getByRole("tab", { name: "Prompt Caching" }).click();
+    await page.getByRole("tab", { name: "By model" }).click();
+    const rows = page.getByRole("row").filter({ hasText: marker });
+    await expect(rows).toHaveCount(1);
+    await expect(rows.first()).toContainText(group);
+    await expect(rows.first()).toContainText("275,000");
+    await expect(
+      page.getByRole("row").filter({ hasText: deployment }),
+    ).toHaveCount(0);
+  } finally {
+    support("clear", apiKey);
+  }
+});

tests/e2e/ui/tests/integrationCritical/expected.json

@@ -6,5 +6,6 @@
   "tests/e2e/ui/tests/integrationCritical/mcpUserEnvVars.spec.ts::pressing Enter on Update opens the credentials modal instead of the server editor",
   "tests/e2e/ui/tests/integrationCritical/mcpUserEnvVars.spec.ts::a server with two per-user variables reports the remaining gap until both are saved",
   "tests/e2e/ui/tests/integrationCritical/mcpUserEnvVars.spec.ts::a server without per-user variables shows no credential row",
-  "tests/e2e/ui/tests/integrationCritical/mcpUserEnvVars.spec.ts::clearing credentials for a server deleted underneath the modal reports the failure without losing the page"
+  "tests/e2e/ui/tests/integrationCritical/mcpUserEnvVars.spec.ts::clearing credentials for a server deleted underneath the modal reports the failure without losing the page",
+  "tests/e2e/ui/tests/integrationCritical/costOptimizationModelGroups.spec.ts::cache leakage by model merges a deployment's resolved and requested model names into its model group"
 ]

tests/integration/_support/daily_spend_rows.py

@@ -0,0 +1,52 @@
+import json
+import sys
+from datetime import datetime, timezone
+from typing import Final, LiteralString
+
+from integration._support.database import read_rows, write_rows
+
+SEED_QUERY: Final[LiteralString] = """
+INSERT INTO "LiteLLM_DailyUserSpend" (
+    id, user_id, date, api_key, model, model_group, custom_llm_provider,
+    endpoint, mcp_namespaced_tool_name,
+    prompt_tokens, completion_tokens, spend,
+    api_requests, successful_requests, failed_requests, updated_at
+) VALUES (
+    gen_random_uuid()::text, %s, %s, %s, %s, %s, 'bedrock',
+    NULL, NULL,
+    %s, %s, %s,
+    %s, %s, %s, now()
+)
+"""
+
+CLEAR_COUNT_QUERY: Final[LiteralString] = 'SELECT count(*) AS count FROM "LiteLLM_DailyUserSpend" WHERE api_key=%s'
+CLEAR_QUERY: Final[LiteralString] = 'DELETE FROM "LiteLLM_DailyUserSpend" WHERE api_key=%s'
+
+
+def seed(api_key: str, group: str, deployment: str) -> int:
+    today: Final = datetime.now(timezone.utc).date().isoformat()
+    write_rows(
+        SEED_QUERY,
+        (api_key, today, api_key, deployment, group, "270000", "1000", "0.81", "270", "270", "0"),
+    )
+    write_rows(
+        SEED_QUERY,
+        (api_key, today, api_key, group, "", "5000", "0", "0.0", "50", "0", "50"),
+    )
+    return 2
+
+
+def clear(api_key: str) -> int:
+    count: Final = int(str(read_rows(CLEAR_COUNT_QUERY, (api_key,))[0]["count"]))
+    write_rows(CLEAR_QUERY, (api_key,))
+    return count
+
+
+if __name__ == "__main__":
+    command: Final = sys.argv[1]
+    if command == "seed":
+        sys.stdout.write(json.dumps({"affected": seed(sys.argv[2], sys.argv[3], sys.argv[4])}) + "\n")
+    elif command == "clear":
+        sys.stdout.write(json.dumps({"affected": clear(sys.argv[2])}) + "\n")
+    else:
+        raise SystemExit(f"unknown command: {command}")

tests/integration/_support/database.py

@@ -1,5 +1,5 @@
 import os
-from typing import Final
+from typing import Final, LiteralString
 
 import psycopg
 from psycopg.rows import dict_row
@@ -14,3 +14,8 @@ def read_rows(
     with psycopg.connect(database_url or os.environ["DATABASE_URL"], row_factory=dict_row) as connection:
         connection.execute("SET TRANSACTION READ ONLY")
         return ROWS.validate_python(connection.execute(query, parameters).fetchall())
+
+
+def write_rows(query: LiteralString, parameters: tuple[str, ...], *, database_url: str | None = None) -> None:
+    with psycopg.connect(database_url or os.environ["DATABASE_URL"]) as connection:
+        connection.execute(query, parameters)

ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/CacheLeakageCard.test.tsx

@@ -46,13 +46,13 @@ const dayWithModels = (date: string, models: Record<string, Partial<SpendMetrics
   date,
   metrics: baseMetrics({}),
   breakdown: {
-    models: Object.fromEntries(
+    models: {},
+    model_groups: Object.fromEntries(
       Object.entries(models).map(([name, m]) => [
         name,
         { metrics: baseMetrics(m), metadata: {}, api_key_breakdown: {} },
       ]),
     ),
-    model_groups: {},
     mcp_servers: {},
     providers: {},
     api_keys: {},

ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/costOptimizationUtils.test.ts

@@ -56,10 +56,10 @@ const modelDay = (date: string, models: Record<string, Partial<SpendMetrics>>):
   date,
   metrics: metrics({}),
   breakdown: {
-    models: Object.fromEntries(
+    models: {},
+    model_groups: Object.fromEntries(
       Object.entries(models).map(([name, m]) => [name, { metrics: metrics(m), metadata: {}, api_key_breakdown: {} }]),
     ),
-    model_groups: {},
     mcp_servers: {},
     providers: {},
     entities: {},
@@ -244,6 +244,54 @@ describe("computeCacheLeakage by model", () => {
     expect(rows.map((r) => r.id)).toEqual(["gemini-2.5-flash"]);
     expect(rows[0].potentialSavings).toBeCloseTo(1.0, 6);
   });
+
+  it("merges rows logged under a deployment's resolved and requested names into one model group row", () => {
+    const day: DailyData = {
+      date: "2026-07-01",
+      metrics: metrics({}),
+      breakdown: {
+        models: {
+          "bedrock/global.anthropic.claude-sonnet-4-6": {
+            metrics: metrics({ prompt_tokens: 270000 }),
+            metadata: {},
+            api_key_breakdown: {},
+          },
+          "bedrock/claude-sonnet-4-6": {
+            metrics: metrics({ prompt_tokens: 5000 }),
+            metadata: {},
+            api_key_breakdown: {},
+          },
+        },
+        model_groups: {
+          "bedrock/claude-sonnet-4-6": {
+            metrics: metrics({ prompt_tokens: 275000 }),
+            metadata: {},
+            api_key_breakdown: {},
+          },
+        },
+        mcp_servers: {},
+        providers: {},
+        entities: {},
+        api_keys: {},
+      },
+    };
+    const { rows } = computeCacheLeakage([day], "model");
+    expect(rows.map((r) => r.id)).toEqual(["bedrock/claude-sonnet-4-6"]);
+    expect(rows[0].uncachedPromptTokens).toBe(275000);
+  });
+
+  it("sums a model group across days", () => {
+    const results = [
+      modelDay("2026-07-01", { "bedrock/claude-sonnet-4-6": { prompt_tokens: 1000 } }),
+      modelDay("2026-07-02", {
+        "bedrock/claude-sonnet-4-6": { prompt_tokens: 2500, cache_read_input_tokens: 500 },
+      }),
+    ];
+    const { rows } = computeCacheLeakage(results, "model");
+    expect(rows).toHaveLength(1);
+    expect(rows[0].uncachedPromptTokens).toBe(3000);
+    expect(rows[0].cacheHitRatio).toBeCloseTo(500 / 3500, 6);
+  });
 });
 
 describe("buildDailyToolSeries", () => {

ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/costOptimizationUtils.ts

@@ -93,7 +93,7 @@ const aggregateByKey = (results: readonly DailyData[]): Map<string, LeakageAccum
 const aggregateByModel = (results: readonly DailyData[]): Map<string, LeakageAccumulator> => {
   const byModel = new Map<string, LeakageAccumulator>();
   for (const day of results) {
-    for (const [model, entry] of Object.entries(day.breakdown?.models ?? {})) {
+    for (const [model, entry] of Object.entries(day.breakdown?.model_groups ?? {})) {
       const acc = byModel.get(model) ?? emptyAccumulator();
       byModel.set(model, addMetrics(acc, entry.metrics, null, null));
     }